You are on page 1of 6

Digital Soil Assessments and Beyond Minasny, Malone & McBratney (eds) 2012 Taylor & Francis Group,

p, London, ISBN 978-0-415-62155-7

Spatial variability of the active layer, permafrost, and soil profile depth in Alaskan soils
U. Mishra & W.J. Riley
Earth Sciences Division, Lawrence Berkeley National Laboratory, US

ABSTRACT: High-latitude carbon-climate-feedback predictions depend on predicted Active-Layer (AL) depth, permafrost (PF) depth, and Whole-Profile (WP) depth in permafrost-affected soils, yet only a few studies have reported these parameters at regional scales. We used spatially referenced soil profile (O-C horizon) description data (n = 472), environmental variables (topographic attributes, climate, land cover, and bedrock geology), and a local regression kriging approach to predict the spatial variability of AL, PF, and WP depths across Alaska at 60 m spatial resolution. We predicted that AL, PF, and WP depths ranged from 2192, 7200, and 18241 cm, respectively. The lowest average prediction error (26 cm) and highest ratio of performance to deviation (1.3) was obtained for AL depth. Our study showed temperature and land cover types as primary controllers of AL depth spatial variability. Elevation and precipitation were the primary controllers of WP depth. Lower predicted spatial variability in comparison to observed data indicate that other spatial datasets than used in this study might improve regional-scale prediction of AL, PF, and WP depths. 1 INTRODUCTION spatial variability at regional and circumpolar scales remains a major scientific challenge (Riseborough et al., 2008). Previous modeling studies have proposed several approaches (e.g., conceptual, empirical, process-based) to estimate the spatial variability of AL depths. The most significant recent advance in these approaches has been Earth System Models for spatial prediction of permafrost extent. However, these studies have in general predicted AL depths in isolation from PF or WP depths, so they do not provide a comprehensive understanding of the portion of total SOC stocks vulnerable to climatic perturbations. Furthermore, the predicted spatial variability at larger scales does not represent the finer-scale variation in environmental parameters. Thus, tracking subgrid heterogeneity (such as the effect of topography) over a regional scale, such that the predicted AL depths represent the heterogeneity of environmental variables, remains a major scientific challenge (Riseborough et al., 2008). In this study, we gathered spatially referenced soil profile (O-C horizon) description data (n = 472) and data on environmental variables (topographic attributes, climate, land cover, and bedrock geology) to predict the spatial variability of AL, PF, and WP depths across Alaska at a fine (60 m) spatial resolution. We then applied a local regression kriging (RK) approach to predict the spatial variability at regional scale, and then compared our results with other published studies

The total carbon stored in permafrost-affected soils is about double that currently in the atmosphere (Schuur et al., 2009). High-latitude regions, where permafrost-affected soils are common are expected to experience much higher temperature increases than temperate or tropical regions over the next century (IPCC, 2007). Since temperature is a primary controller of decomposition, these soils are a potentially vulnerable component of the global carbon cycle. Permafrost-affected (PF) soil is defined as soil in which some portion of the profile remains at or below 0C for at least two consecutive years. Within a PF soil, the active-layer (AL) is defined as the top portion that freezes and thaws annually, and the whole-profile (WP) is defined as the sum of AL and PF depths (i.e., from O to C horizons). Ground-temperature observations from Alaska over the last 30 years have indicated an increase in permafrost temperatures between 0.5 and 3C (Osterkemp, 2005). These temperature increases have been linked to increased AL depths and decreased PF extent. Significant increases in temperature (3.23.5C) that are expected by 2100 in high latitude regions as a result of future climate change (IPCC, 2007) could deepen the AL depth, decreasing the permafrost thickness and moving more soil organic carbon (SOC) into the active layer. Despite recent advancements in predicting these dynamics, prediction of AL and PF depth

83

that used both observational and process-based modeling approaches. 2 2.1 MATERIALS AND METHODS Study area and soil profile observations

This study was conducted in the state of Alaska, USA, covering a land area of 1,221,272 km2, excluding the surface areas of ice, water, and bare rocks. A total of 422 geo-referenced SOC profile data were collected from the National Soil Survey Characterization database (National Soil Survey Laboratory, 2010). This soil survey database contained measurements from representative soil profiles throughout Alaska, covering all soil types at the soil suborder level (18 suborders). To this collection, we added an additional 50 soil profile observations from the Arctic regions of North America (Ping et al., 2008). Though the SOC profile samples were unevenly distributed throughout the study area (Fig. 1), the samples covered all 27 major land resource areas (MLRA) in the state (The MLRA is a geographical unit that contains similar patterns of climate, soils, water resources, and land uses). The presence of a permafrost layer was indicated by using a genetic horizon f (i.e., frozen layer) in the soil profile data. The average depth to the f horizon was assumed to be equal to the AL depth of the soil profile. 2.2 Environmental datasets

area, and low-elevation areas in the western and northern parts. After we extracted land-cover data (of 60 m spatial resolution) for Alaska from the NLCD database (Multi-Resolution Land Characteristics Consortium, 2011), we reclassified the NLCD land-cover types into nine major categories (Table 1): scrub (the largest category at 43%), forest (25%), barren (8.5%), herbaceous (7%), and wetlands (7%). The remaining surface area (9.5%) was under open water, perennial ice, barren land, and moss vegetation. Indicator variables for the presence or absence of seven land-cover types (except open water and perennial ice) were created and used in the model selection process. Climate data, such as long-term (19611990) mean annual air temperature and mean annual precipitation, were obtained from the PRISM database of spatial climate analysis at Oregon State University (Daly et al., 2001). The mean annual air temperature and mean annual precipitation ranged from 18C to 6C and from 150 mm to 8500 mm, respectively. Both temperature and precipitation increased from north to south. The bedrock geology data, obtained from a USGS database, indicated 180 types of bedrock across Alaska. The largest land area was under Quaternary rock deposits (8%), followed by Cretaceous rock (7.3%), Lower Paleozoic rock (6.6%), Lower Cretaceous rock (6.2%), ice (4.3%), and Pleistocene deposits (4.2%). The remaining surface area was under 174 bedrock types. 2.3 Spatial modeling and accuracy of predictions

We obtained a digital elevation model of 60 m spatial resolution from the USGS database (Multi-Resolution Land Characteristics Consortium, 2011). Study-area elevation ranged from sea level to 6188 m, with high-elevation areas located in the eastern and southern parts of the study

With these data, we applied a local regression kriging approach to predict AL, PF, and WP depths.
Table 1. Reclassification of USGS land-cover types for this study. NLCD land cover type Developed open space, low intensity, medium intensity and high intensity Deciduous, evergreen, and mixed forest Dwarf scrub and shrub scrub Shrub, sedge, and moss Pasture and cultivated lands Woody and herbaceous wetlands Barren Open water Perennial ice Reclassified land cover type Developed Percent land cover type 0.09

Forest Scrub Herbaceous Cultivated Wetland Barren Open water Perennial ice

25.1 43.1 7.1 0.02 7.1 8.5 4.0 4.7

Figure 1. Distribution of soil organic carbon profile observations across Alaska.

84

For this purpose, we adopted a Mallows Cp criteria for model selection, in a best subset regression approach, to identify the environmental variables that could be used in further modeling steps. Selected independent variables were tested for unequal error variance, multicollinearity of variables, normality, and randomness of residuals. Selected independent variables were then used in a geographically weighted regression (GWR) approach, and the model parameters were derived at a 1,000 m regular interval throughout the study area. The residuals obtained from this method both for whole profile depth and active layer depth showed spatial autocorrelation (Table 2) and were interpolated using a moving-window ordinary kriging approach, in which variograms were recalculated at the same interval across the study area. Geostatistical Analyst tool of ArcGIS (10) was used for this purpose. The local RK process adopted in this study can be summarized as: DLRK ( x ) = DGWR ( x ) + MOK ( x )
^ ^ ^ ^

(1)

DLRK ( x ) = predicted values of where the depth using local regression kriging; ^ DGWR ( x ) = geographically weighted re regression e W predicted values of the variable; and MOK ( x ) is the kriged values of the residuals at point x using a moving window ordinary kriging. In GWR, the weight function is chosen as an adaptive spatial kernel type, so that the spatial extent for samples included variations based on sample density. The bandwidth was chosen based on Akaike Information Criterion minimization, and the 21 samples were included. The GWR approach used in this study can be represented as:
^

the difference between WP and AL depth. We note that the reported AL depths may not have been the maximum AL depth of the year, and that these measurements occurred over a period of 50 (1952 2002) years with most of the profile observations made between 19601990. We expect error is introduced to our extrapolation because of inter-annual variability in climate that can drive variability in AL depths. In permafrost soils, typical temporal profiles of AL depth begin with a rapid deepening immediately after snow melt, followed by a relatively stable profile over the snow-free period, and then a rapid return to the surface during winter. The prediction accuracy of the resulting depth maps was evaluated using a K-fold validation approach (Martin et al., 2011). For this purpose, the entire dataset was divided into calibration (n = 412) and validation (n = 60) datasets five times. Mapping of depth using calibration datasets and their validation was conducted for each split, and the average validation indices are reported here. From the predicted depth maps, depth values were extracted for the validation points. The obtained values of the observed and predicted depths were interpreted by calculating the mean estimation error (MEE) and root mean square error (RMSE): MEE = E
^ 1 n ( D( xi ) D( xi )) n i =1 ^ 1 n ( D( xi ) D( xi ))2 n i =1 ^

(3)

RMSE =

(4)

Di

0 (ui vi ) 1(ui ,vi )X i1 (u


+ 2 (ui ,vi )X i 2 (u X
^ ^

k (ui ,vi )X ik

where Di = predicted depth at location i ( i ,vi ) are ,v the coordinates for location i; 0 to = regression coefficients; Xi1 to Xik = environmental variables at location i; and k is the number of environmental variables. We applied the above approach for AL and WP depths separately. The PF depth was considered as
Table 2. Morans index (I) statistic showing spatial autocorrelation in depth residuals. Variables Residual profile depth Residual active layer depth Morans I 0.50 0.15 P value 0.01 0.05

where D(xi ) = measured depth; D( xi ) = estimated depth; and n = number of validated observations. These values should approach zero for an optimal prediction. We also calculated the ratio of performance to deviation (RPD, defined as the ratio between the standard deviation and the RMSE), which indicates the overall prediction ability of the selected approach. 3 3.1 RESULTS AND DISCUSSION Descriptive statistics

Across the 472 pedons, observed AL, PF, and WP depths had a large range with unimodal and positively skewed distributions (Table 3). The average observed AL depth for all of Alaska was 47 cm, ranging from 2 to 250 cm, and showing a high spatial variability (coefficient of variability, CV = 61%). The average observed PF depth was 106 cm, ranging from 7 to 400 cm and showing a comparably high spatial variability (CV = 62%). The average observed whole-profile depth was

85

Table 3. Descriptive statistics of observed soil properties (total dataset, n = 472; permafrost profiles, n = 133). Parameters Min Max Ave Med CV (%) Skew SD 450 116 100 250 46.7 42 408 106 74 54.4 62 62 1.4 3.2 1.1 63.3 28.8 83.6

Whole-profile 10 depth (cm) Active-layer 2 depth (cm) Permafrost 7 depth (cm)

* Min = minimum; Max = maximum; Ave = average; Med = median; CV = coefficient of variability; Skew = skewness; and SD = standard deviation.

116 cm, ranging from 10 to 450 cm and with a CV of 54%. The observations indicated deeper AL depths in the southern parts and deeper PF depths in the northern and eastern parts of the state. 3.2 Spatial variability of predicted active-layer, permafrost, and whole-profile depths

Using the geospatial extrapolation approach described above, we predicted an average AL depth for Alaska of 37 cm, with high spatial variability (CV = 52%) and a range of 2 to 192 cm. Over most of the state, the predicted AL depth was less than 40 cm; higher predicted AL depths had sporadic spatial distribution (Fig. 2(a)). The average error of AL depth prediction (RMSE) was 26.5 cm, with a ratio of performance to deviation (RPD) of 1.3. The average predicted PF depth for Alaska was 68 cm, ranging from 7 to 200 cm. In contrast to the observations, predicted spatial variability was moderate (CV = 29%); the RMSE was 82 cm and RPD was 0.97. The average whole-profile depth of the state was predicted to be 105 cm, ranging from 18 to 241 cm. Predicted spatial variability was relatively lower (CV = 16.5%) than for either of the two component depth intervals or for the observations. The average prediction error for WP depth was 60 cm, with an RPD of 1. The validation indices showed highest prediction accuracy for AL depth prediction and lowest accuracy for PF depth predictions across the state (Table 4). We compared our results with observed and previously modeled results. Bockheim (2007) used 21 pedons across Alaska that were not included in the dataset applied here and reported an average AL depth of 47.5 cm (3090 cm). Michaelson et al. (1996), using a different set of 25 pedons from Northern Alaska, reported an average AL depth of 51 cm (26 >100 cm). Our predicted mean and range of AL depth are smaller and larger, respectively, than these observation-based studies. Previous numerical model predictions of AL depths

Figure 2. Predicted active-layer depth (a), permafrost depth (b), and whole-profile depth (c) of Alaskan soils.

Table 4. Validation indices for active-layer, permafrost, and whole-profile depths. Parameters Active-layer Permafrost Whole-profile 24 82 0.97 14 60 1

MEE (cm) 2 RMSE (cm) 26 RPD 1.2

* MEE = mean estimation error; RMSE = root mean square error; and RPD = ratio of performance to deviation.

were deeper than our observed and predicted values. For instance, Marchenko et al. (2008), using a spatially-distributed numerical modeling approach, predicted AL depth across Alaska to be

86

0300 cm. Schaefer et al. (2011), using a climatescale land-surface model, predicted AL depths across the entire northern circumpolar region; their predictions for Alaska ranged from 80 to 200 cm. Our predictions are within the lower range of these large-scale modeling predictions (2192 cm). We are unaware of published numerical modeling or observational studies predicting PF and WP depths. 3.3 Controls of active-layer, permafrost, and whole-profile depths

(Nelson et al., 1998). Both WP and AL depth decreased with increased precipitation. However, no trend with precipitation was observed for PF depths. 4 SUMMARY AND LIMITATIONS

Largest predicted AL depths and smallest PF depths were found under pasture and barren landcover types, respectively. Smallest AL depths and largest PF depths were found under scrub and herbaceous vegetations. Forest and wetlands had similar AL and PF depths that were intermediate between these two groups of land cover types. Largest WP depths were predicted under herbaceous and wetland land covers, while smallest WP depths were found under barren land cover. Predicted WP and PF depths decreased with increasing elevation. In contrast, predicted AL depths were relatively constant over various elevation ranges. The negative relationship between WP depth and elevation can be explained by considering the soil profile thickness as a balance between production and erosion of soil (Heimsath et al., 1997). However, AL depths also depend on microclimate, soil moisture, vegetation, and topographic features (Nelson et al., 1998). The AL depth increased and PF depth decreased with increasing temperature, but no trend with temperature was observed for WP depth (Fig. 3). This trend of AL variability is mainly because the AL depth varies both spatially and temporally, in response to air temperature. As a result, mean annual air temperature has been used by various authors to model the AL thickness
Whole-prole Depth Permafrost Depth 140 120 100 80 60 40 20 0 Ac ve-layer Depth

In this study we used soil profile observations and a geospatial interpolation approach to predict the spatial distribution of AL, PF, and WP depths at a fine spatial resolution across Alaska. The higher prediction accuracy of AL depths compared to PF and WP depths illustrates the greater influence of environmental variables in determining AL distribution. Our results indicate that two recent process-based modeling studies overestimated the AL depths of high-latitude soils. The dependence of our predicted depths was consistent with existing theory. Lower predicted spatial variability compared to observations indicate that the spatial datasets for other soil-forming factors that are important for high-latitude environments, such as time since pedogenesis, fire frequency, and fire intensity, could increase the variability of AL, PF, and WP depth predictions. Likewise, kriging tends to smooth the variability as it is an interpolation technique. The data density was not adequate for robust detection of trends at all scales and prediction accuracy of our results could have been increased by including more homogeneously-distributed SOC profile samples across Alaska. Furthermore, the PF depth was not independent as it was calculated as the difference between WP and AL depths. REFERENCES
Bockheim, J.G. 2007. Importance of cryoturbation in redistributing organic carbon in permafrost-affected soils. Soil Science Society of America Journal 71:13351342. Daly, C., Taylor, G.H., Gibson, W.P., Parzybok, T.W., Johnson, G.L. & Pasteris, P. 2001. High quality spatial climate data sets for the United States and beyond. Transactions of the American Society of Agricultural Engineers 43:19571962. Fotheringham, A.S., Brunsdon, C. & Charlton, M. 2002. Geographically weighted regression: The analysis of spatially varying relationships. John Wiley & Sons Ltd., Chichester, UK. Heimsath, A.M., Dietrich, W.E., Nishiizumi, K. & Finkel, R.C. 1997. The soil production function and landscape equilibrium. Nature 388:358361. IPCC 2007. Summary for policy makers. In: Climate change 2007: The physical science basis. Contribution of Working Group 1 to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change, Cambridge University Press, New York.

Depth (cm)

< -12 -12 - -8 -8 - - 4 -4 - 0 (n=103) (n=20) (n=109) (n=28)

0-4 4-6 (n=67) (n=145)

Average Annual Temperature (oC)

Figure 3. Temperature control on whole-profile, active-layer, and permafrost depths. Error bars represent 1 standard deviation.

87

Marchenko, S., Romanovsky, V. & Tipenko, G. 2008. Numerical modeling of spatial permafrost dynamics in Alaska. Proceedings Ninth International Conference on Permafrost 2:11251130. Martin, M.P., Wattenbach, M., Smith, P., Meersmans, J., Jolivet, C.L., Boulonne, L. & Arrouays, D. 2011. Spatial distribution of soil organic carbon stocks in France. Biogeosciences 8:10531065. Michaelson, G.J., Ping, C.L. & Kimble, J.M. 1996. Carbon storage and distribution in tundra soils of arctic Alaska, U.S.A. Arctic & Alpine Research 28:414424. Multi-Resolution Land Characteristics Consortium, 2007. 2001 National land cover data (NLCD 2001). Available at www.epa.gov/mrlc/nlcd2001.html (Accessed 20 Feb. 2011). USEPA, Washington, DC. National Soil Survey Laboratory, 2011. National cooperative soil characterization database. Available at ssldata.nrcs.usda.gov/(Accessed 20 Mar. 2011). Lincoln, NE.

Nelson, F.E., Hinkel, K.M., Shiklomanov, N.I., Mueller, G.R., Miller, L.L. & Walker, D.A. 1998. Active-layer thickness in north central Aalska: Systematic sampling, scale, and spatial autocorrelation. Journal of Geophysical Research 103:2893628973. Osterkamp, T.E. 2005. The recent warming of permafrost in Alaska. Global and Planetary Change 49:187202. Riseborough, D., Shiklomanov, N., Etzelmuller, B., Gruber, S. & Marchenko, S. 2008. Recent Advances in Permafrost Modelling. Permafrost and Periglacial Processes 19:137156. Schaefer, K., Zhang, T., Bruhwiler, L. & Barrett, A.P. 2011. Amount and timing of permafrost carbon release in response to climate warming. Tellus 63B:165180. Schuur, E.A.G., Vogel, J.G., Crummer, K.G., Lee, H., Sickman, J.O. & Osterkamp, T.E. 2009 The effect of permafrost thaw on old carbon release and net carbon exchange from tundra. Nature 459:556559.

88

You might also like